Maki Postdoctoral Associate Desert Research Institute
Process-based models such as Variable Infiltration Capacity (VIC) have been used to create streamflow projections in response to climate change. These models are typically calibrated against historical observations in the watershed where streamflow projections are required. A problem is that the calibrated parameters of these models depend upon the climate: climate change may result in a change in the model parameters. Thus, a process-based model may yield unreliable projections of streamflow if the model is calibrated on historical data only in the watershed where projections are required. One way to deal with this problem is to use historical data across several watersheds to extract information about the streamflow projections. This is the so-called ‘space-time symmetry’ idea. This idea is implemented using a machine learning algorithm called Long-Short Memory Network (LSTM).
The aim of this study is twofold: (1) To test the ability of LSTMs to extract streamflow information from data across several watersheds in the presence of climate change, and (2) To check the physical realism of streamflow projections developed by the LSTM model. Toward this end, several LSTM models were trained using historical data from the different numbers of watersheds varying from 1 to 461. Subsequently, the optimal number of training watersheds was identified based on the performance of these models. This information was subsequently used to create streamflow projections across the USA which were compared with the projections obtained by using the VIC model. In this presentation, the detailed results of this study will be presented.
